[PATCH] D80432: AMDGPU: Refine rcp/rsq intrinsic folding for modern FP rules

[Matt Arsenault via Phabricator via llvm-commits]

arsenm added a comment.

In D80432#2050967 <https://reviews.llvm.org/D80432#2050967>, @foad wrote:

> > We have to assume undef could be an snan, which would need quieting so returning qnan is safer than undef.
>
> So you've chosen to optimize assuming that an undef input was a nan. Is that better than assuming it was something more ordinary like 0?


qnan matches the current fdiv handling. nan also enables folding out more use operations more consistently


CHANGES SINCE LAST ACTION
  https://reviews.llvm.org/D80432/new/

https://reviews.llvm.org/D80432